xref: /titanic_44/usr/src/uts/common/xen/dtrace/xdt.c (revision 444ce08e035c2cafaa89f7236e38bbe82a287904)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright 2008 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 #pragma ident	"%Z%%M%	%I%	%E% SMI"
28 
29 /*
30  * Xen event provider for DTrace
31  *
32  * NOTE: This provider is PRIVATE. It is intended as a short-term solution and
33  * may disappear or be re-implemented at anytime.
34  *
35  * This provider isn't suitable as a general-purpose solution for a number of
36  * reasons. First and foremost, we rely on the Xen tracing mechanism and don't
37  * have any way to gather data other than that collected by the Xen trace
38  * buffers. Further, it does not fit into the DTrace model (see "Interacting
39  * with DTrace" below.)
40  *
41  *
42  * Tracing in Xen
43  * --------------
44  *
45  * Xen implements a tracing facility for generating and collecting execution
46  * event traces from the hypervisor. When tracing is enabled, compiled in
47  * probes record events in contiguous per-CPU trace buffers.
48  *
49  *               +---------+
50  * +------+      |         |
51  * | CPUn |----> | BUFFERn |
52  * +------+      |         |
53  *               +---------+- tbuf.va + (tbuf.size * n)
54  *               :         :
55  *               +---------+
56  * +------+      |         |
57  * | CPU1 |----> | BUFFER1 |
58  * +------+      |         |
59  *               +---------+- tbuf.va + tbuf.size
60  * +------+      |         |
61  * | CPU0 |----> | BUFFER0 |
62  * +------+      |         |
63  *               +---------+- tbuf.va
64  *
65  * Each CPU buffer consists of a metadata header followed by the trace records.
66  * The metadata consists of a producer/consumer pair of pointers into the buffer
67  * that point to the next record to be written and the next record to be read
68  * respectively. The trace record format is as follows:
69  *
70  * +--------------------------------------------------------------------------+
71  * | CPUID(uint_t) | TSC(uint64_t) | EVENTID(uint32_t) |     DATA FIELDS      |
72  * +--------------------------------------------------------------------------+
73  *
74  * DATA FIELDS:
75  * +--------------------------------------------------------------------------+
76  * | D1(uint32_t) | D2(uint32_t) | D3(uint32_t) | D4(uint32_t) | D5(uint32_t) |
77  * +--------------------------------------------------------------------------+
78  *
79  *
80  * Interacting with DTrace
81  * -----------------------
82  *
83  * Every xdt_poll_nsec nano-seconds we poll the trace buffers for data and feed
84  * each entry into dtrace_probe() with the corresponding probe ID for the event.
85  * As a result of this periodic collection implementation probe firings are
86  * asynchronous. This is the only sensible way to implement this form of
87  * provider, but because of its asynchronous nature asking things like
88  * "current CPU" and, more importantly, arbitrary questions about the context
89  * surrounding the probe firing are not meaningful. So, consumers should not
90  * attempt to infer anything beyond what is supplied via the probe arguments.
91  */
92 
93 #include <sys/types.h>
94 #include <sys/sysmacros.h>
95 #include <sys/modctl.h>
96 #include <sys/sunddi.h>
97 #include <sys/ddi.h>
98 #include <sys/conf.h>
99 #include <sys/devops.h>
100 #include <sys/stat.h>
101 #include <sys/cmn_err.h>
102 #include <sys/dtrace.h>
103 #include <sys/sdt.h>
104 #include <sys/cyclic.h>
105 #include <vm/seg_kmem.h>
106 #include <vm/hat_i86.h>
107 #include <sys/hypervisor.h>
108 #include <xen/public/trace.h>
109 #include <xen/public/sched.h>
110 
111 #define	XDT_POLL_DEFAULT	100000000	/* default poll interval (ns) */
112 #define	XDT_POLL_MIN		10000000	/* min poll interval (ns) */
113 #define	XDT_TBUF_RETRY		50		/* tbuf disable retry count */
114 
115 /*
116  * The domid must match IDLE_DOMAIN_ID in xen.hg/xen/include/xen/sched.h
117  * in the xVM gate.
118  */
119 #define	IS_IDLE_DOM(domid)	(domid == 0x7FFFU)
120 
121 /* Macros to extract the domid and cpuid from a HVM trace data field */
122 #define	HVM_DOMID(d)		(d >> 16)
123 #define	HVM_VCPUID(d)		(d & 0xFFFF)
124 
125 #define	XDT_PROBE4(event, cpuid, arg0, arg1, arg2, arg3) {		\
126 	dtrace_id_t id = xdt_probemap[event];				\
127 	if (id)								\
128 		dtrace_probe(id, cpuid, arg0, arg1, arg2, arg3);	\
129 }									\
130 
131 #define	XDT_PROBE3(event, cpuid, arg0, arg1, arg2) \
132 	XDT_PROBE4(event, cpuid, arg0, arg1, arg2, 0)
133 
134 #define	XDT_PROBE2(event, cpuid, arg0, arg1) \
135 	XDT_PROBE4(event, cpuid, arg0, arg1, 0, 0)
136 
137 #define	XDT_PROBE1(event, cpuid, arg0) \
138 	XDT_PROBE4(event, cpuid, arg0, 0, 0, 0)
139 
140 #define	XDT_PROBE0(event, cpuid) \
141 	XDT_PROBE4(event, cpuid, 0, 0, 0, 0)
142 
143 /* Probe classes */
144 #define	XDT_SCHED			0
145 #define	XDT_MEM				1
146 #define	XDT_HVM				2
147 #define	XDT_NCLASSES			3
148 
149 /* Probe events */
150 #define	XDT_EVT_INVALID			(-(int)1)
151 #define	XDT_SCHED_OFF_CPU		0
152 #define	XDT_SCHED_ON_CPU		1
153 #define	XDT_SCHED_IDLE_OFF_CPU		2
154 #define	XDT_SCHED_IDLE_ON_CPU		3
155 #define	XDT_SCHED_BLOCK			4
156 #define	XDT_SCHED_SLEEP			5
157 #define	XDT_SCHED_WAKE			6
158 #define	XDT_SCHED_YIELD			7
159 #define	XDT_SCHED_SHUTDOWN_POWEROFF	8
160 #define	XDT_SCHED_SHUTDOWN_REBOOT	9
161 #define	XDT_SCHED_SHUTDOWN_SUSPEND	10
162 #define	XDT_SCHED_SHUTDOWN_CRASH	11
163 #define	XDT_MEM_PAGE_GRANT_MAP		12
164 #define	XDT_MEM_PAGE_GRANT_UNMAP	13
165 #define	XDT_MEM_PAGE_GRANT_TRANSFER	14
166 #define	XDT_HVM_VMENTRY			15
167 #define	XDT_HVM_VMEXIT			16
168 #define	XDT_NEVENTS			17
169 
170 typedef struct {
171 	const char	*pr_mod;	/* probe module */
172 	const char	*pr_name;	/* probe name */
173 	int		evt_id;		/* event id */
174 	uint_t		class;		/* probe class */
175 } xdt_probe_t;
176 
177 typedef struct {
178 	uint32_t	trc_mask;	/* trace mask */
179 	uint32_t	cnt;		/* num enabled probes in class */
180 } xdt_classinfo_t;
181 
182 typedef struct {
183 	ulong_t prev_domid;		/* previous dom executed */
184 	ulong_t prev_vcpuid;		/* previous vcpu executed */
185 	ulong_t prev_ctime;		/* time spent on cpu */
186 	ulong_t next_domid;		/* next dom to be scheduled */
187 	ulong_t next_vcpuid;		/* next vcpu to be scheduled */
188 	ulong_t next_wtime;		/* time spent waiting to get on cpu */
189 	ulong_t next_ts;		/* allocated time slice */
190 } xdt_schedinfo_t;
191 
192 static struct {
193 	uint_t cnt;			/* total num of trace buffers */
194 	size_t size;			/* size of each cpu buffer */
195 	mfn_t start_mfn;		/* starting mfn of buffers */
196 	caddr_t va;			/* va buffers are mapped into */
197 
198 	/* per-cpu buffers */
199 	struct t_buf **meta;		/* buffer metadata */
200 	struct t_rec **data;		/* buffer data records */
201 
202 	/* statistics */
203 	uint64_t stat_dropped_recs;	/* records dropped */
204 	uint64_t stat_spurious_cpu;	/* recs with garbage cpuids */
205 	uint64_t stat_spurious_switch;	/* inconsistent vcpu switches */
206 	uint64_t stat_unknown_shutdown;	/* unknown shutdown code */
207 	uint64_t stat_unknown_recs;	/* unknown records */
208 } tbuf;
209 
210 static char *xdt_stats[] = {
211 	"dropped_recs",
212 };
213 
214 /*
215  * Tunable variables
216  *
217  * The following may be tuned by adding a line to /etc/system that
218  * includes both the name of the module ("xdt") and the name of the variable.
219  * For example:
220  *     set xdt:xdt_tbuf_pages = 40
221  */
222 uint_t xdt_tbuf_pages = 20;			/* pages to alloc per-cpu buf */
223 
224 /*
225  * The following may be tuned by adding a line to
226  * /platform/i86xpv/kernel/drv/xdt.conf.
227  * For example:
228  *     xdt_poll_nsec = 200000000;
229  */
230 static hrtime_t xdt_poll_nsec;			/* trace buffer poll interval */
231 
232 /*
233  * Internal variables
234  */
235 static dev_info_t *xdt_devi;
236 static dtrace_provider_id_t xdt_id;
237 static uint_t xdt_ncpus;			/* total number of phys CPUs */
238 static uint32_t cur_trace_mask;			/* current trace mask */
239 static xdt_schedinfo_t *xdt_cpu_schedinfo;	/* per-cpu sched info */
240 dtrace_id_t xdt_probemap[XDT_NEVENTS];		/* map of enabled probes */
241 dtrace_id_t xdt_prid[XDT_NEVENTS];		/* IDs of registered events */
242 static cyclic_id_t xdt_cyclic = CYCLIC_NONE;
243 static kstat_t *xdt_kstats;
244 static xdt_classinfo_t xdt_classinfo[XDT_NCLASSES];
245 
246 static xdt_probe_t xdt_probe[] = {
247 	/* Sched probes */
248 	{ "sched", "off-cpu", XDT_SCHED_OFF_CPU, XDT_SCHED },
249 	{ "sched", "on-cpu", XDT_SCHED_ON_CPU, XDT_SCHED },
250 	{ "sched", "idle-off-cpu", XDT_SCHED_IDLE_OFF_CPU, XDT_SCHED },
251 	{ "sched", "idle-on-cpu", XDT_SCHED_IDLE_ON_CPU, XDT_SCHED },
252 	{ "sched", "block", XDT_SCHED_BLOCK, XDT_SCHED },
253 	{ "sched", "sleep", XDT_SCHED_SLEEP, XDT_SCHED },
254 	{ "sched", "wake", XDT_SCHED_WAKE, XDT_SCHED },
255 	{ "sched", "yield", XDT_SCHED_YIELD, XDT_SCHED },
256 	{ "sched", "shutdown-poweroff", XDT_SCHED_SHUTDOWN_POWEROFF,
257 		XDT_SCHED },
258 	{ "sched", "shutdown-reboot", XDT_SCHED_SHUTDOWN_REBOOT, XDT_SCHED },
259 	{ "sched", "shutdown-suspend", XDT_SCHED_SHUTDOWN_SUSPEND, XDT_SCHED },
260 	{ "sched", "shutdown-crash", XDT_SCHED_SHUTDOWN_CRASH, XDT_SCHED },
261 
262 	/* Memory probes */
263 	{ "mem", "page-grant-map", XDT_MEM_PAGE_GRANT_MAP, XDT_MEM },
264 	{ "mem", "page-grant-unmap", XDT_MEM_PAGE_GRANT_UNMAP, XDT_MEM },
265 	{ "mem", "page-grant-transfer", XDT_MEM_PAGE_GRANT_TRANSFER, XDT_MEM },
266 
267 	/* HVM probes */
268 	{ "hvm", "vmentry", XDT_HVM_VMENTRY, XDT_HVM },
269 	{ "hvm", "vmexit", XDT_HVM_VMEXIT, XDT_HVM },
270 
271 	{ NULL }
272 };
273 
274 extern uint_t xen_get_nphyscpus(void);
275 
276 static inline uint32_t
277 xdt_nr_active_probes()
278 {
279 	int i;
280 	uint32_t tot = 0;
281 
282 	for (i = 0; i < XDT_NCLASSES; i++)
283 		tot += xdt_classinfo[i].cnt;
284 
285 	return (tot);
286 }
287 
288 static void
289 xdt_init_trace_masks(void)
290 {
291 	xdt_classinfo[XDT_SCHED].trc_mask = TRC_SCHED;
292 	xdt_classinfo[XDT_MEM].trc_mask = TRC_MEM;
293 	xdt_classinfo[XDT_HVM].trc_mask = TRC_HVM;
294 }
295 
296 static int
297 xdt_kstat_update(kstat_t *ksp, int flag)
298 {
299 	kstat_named_t *knp;
300 
301 	if (flag != KSTAT_READ)
302 		return (EACCES);
303 
304 	knp = ksp->ks_data;
305 
306 	/*
307 	 * Assignment order should match that of the names in
308 	 * xdt_stats.
309 	 */
310 	(knp++)->value.ui64 = tbuf.stat_dropped_recs;
311 
312 	return (0);
313 }
314 
315 static void
316 xdt_kstat_init(void)
317 {
318 	int nstats = sizeof (xdt_stats) / sizeof (xdt_stats[0]);
319 	char **cp = xdt_stats;
320 	kstat_named_t *knp;
321 
322 	if ((xdt_kstats = kstat_create("xdt", 0, "trace_statistics", "misc",
323 	    KSTAT_TYPE_NAMED, nstats, 0)) == NULL)
324 		return;
325 
326 	xdt_kstats->ks_update = xdt_kstat_update;
327 
328 	knp = xdt_kstats->ks_data;
329 	while (nstats > 0) {
330 		kstat_named_init(knp, *cp, KSTAT_DATA_UINT64);
331 		knp++;
332 		cp++;
333 		nstats--;
334 	}
335 
336 	kstat_install(xdt_kstats);
337 }
338 
339 static int
340 xdt_sysctl_tbuf(xen_sysctl_tbuf_op_t *tbuf_op)
341 {
342 	xen_sysctl_t op;
343 	int xerr;
344 
345 	op.cmd = XEN_SYSCTL_tbuf_op;
346 	op.interface_version = XEN_SYSCTL_INTERFACE_VERSION;
347 	op.u.tbuf_op = *tbuf_op;
348 
349 	if ((xerr = HYPERVISOR_sysctl(&op)) != 0)
350 		return (xen_xlate_errcode(xerr));
351 
352 	*tbuf_op = op.u.tbuf_op;
353 	return (0);
354 }
355 
356 static int
357 xdt_map_trace_buffers(mfn_t mfn, caddr_t va, size_t len)
358 {
359 	x86pte_t pte;
360 	caddr_t const sva = va;
361 	caddr_t const eva = va + len;
362 	int xerr;
363 
364 	ASSERT(mfn != MFN_INVALID);
365 	ASSERT(va != NULL);
366 	ASSERT(IS_PAGEALIGNED(len));
367 
368 	for (; va < eva; va += MMU_PAGESIZE) {
369 		/*
370 		 * Ask the HAT to load a throwaway mapping to page zero, then
371 		 * overwrite it with the hypervisor mapping. It gets removed
372 		 * later via hat_unload().
373 		 */
374 		hat_devload(kas.a_hat, va, MMU_PAGESIZE, (pfn_t)0,
375 		    PROT_READ | HAT_UNORDERED_OK,
376 		    HAT_LOAD_NOCONSIST | HAT_LOAD);
377 
378 		pte = mmu_ptob((x86pte_t)mfn) | PT_VALID | PT_USER
379 		    | PT_FOREIGN | PT_WRITABLE;
380 
381 		xerr = HYPERVISOR_update_va_mapping_otherdomain((ulong_t)va,
382 		    pte, UVMF_INVLPG | UVMF_LOCAL, DOMID_XEN);
383 
384 		if (xerr != 0) {
385 			/* unmap pages loaded so far */
386 			size_t ulen = (uintptr_t)(va + MMU_PAGESIZE) -
387 			    (uintptr_t)sva;
388 			hat_unload(kas.a_hat, sva, ulen, HAT_UNLOAD_UNMAP);
389 			return (xen_xlate_errcode(xerr));
390 		}
391 
392 		mfn++;
393 	}
394 
395 	return (0);
396 }
397 
398 static int
399 xdt_attach_trace_buffers(void)
400 {
401 	xen_sysctl_tbuf_op_t tbuf_op;
402 	size_t len;
403 	int err;
404 	uint_t i;
405 
406 	/* set trace buffer size */
407 	tbuf_op.cmd  = XEN_SYSCTL_TBUFOP_set_size;
408 	tbuf_op.size = xdt_tbuf_pages;
409 	(void) xdt_sysctl_tbuf(&tbuf_op);
410 
411 	/* get trace buffer info */
412 	tbuf_op.cmd  = XEN_SYSCTL_TBUFOP_get_info;
413 	if ((err = xdt_sysctl_tbuf(&tbuf_op)) != 0)
414 		return (err);
415 
416 	tbuf.size = tbuf_op.size;
417 	tbuf.start_mfn = (mfn_t)tbuf_op.buffer_mfn;
418 	tbuf.cnt = xdt_ncpus;
419 
420 	if (tbuf.size == 0) {
421 		cmn_err(CE_NOTE, "No trace buffers allocated!");
422 		return (ENOBUFS);
423 	}
424 
425 	ASSERT(tbuf.start_mfn != MFN_INVALID);
426 	ASSERT(tbuf.cnt > 0);
427 
428 	len = tbuf.size * tbuf.cnt;
429 	tbuf.va = vmem_alloc(heap_arena, len, VM_SLEEP);
430 
431 	if ((err = xdt_map_trace_buffers(tbuf.start_mfn, tbuf.va, len)) != 0) {
432 		vmem_free(heap_arena, tbuf.va, len);
433 		tbuf.va = NULL;
434 		return (err);
435 	}
436 
437 	tbuf.meta = (struct t_buf **)kmem_alloc(tbuf.cnt * sizeof (*tbuf.meta),
438 	    KM_SLEEP);
439 	tbuf.data = (struct t_rec **)kmem_alloc(tbuf.cnt * sizeof (*tbuf.data),
440 	    KM_SLEEP);
441 
442 	for (i = 0; i < tbuf.cnt; i++) {
443 		void *cpu_buf = (void *)(tbuf.va + (tbuf.size * i));
444 		tbuf.meta[i] = cpu_buf;
445 		tbuf.data[i] = (struct t_rec *)((uintptr_t)cpu_buf +
446 		    sizeof (struct t_buf));
447 
448 		/* throw away stale trace records */
449 		tbuf.meta[i]->cons = tbuf.meta[i]->prod;
450 	}
451 
452 	return (0);
453 }
454 
455 static void
456 xdt_detach_trace_buffers(void)
457 {
458 	size_t len = tbuf.size * tbuf.cnt;
459 
460 	ASSERT(tbuf.va != NULL);
461 
462 	hat_unload(kas.a_hat, tbuf.va, len,
463 	    HAT_UNLOAD_UNMAP | HAT_UNLOAD_UNLOCK);
464 	vmem_free(heap_arena, tbuf.va, len);
465 	kmem_free(tbuf.meta, tbuf.cnt * sizeof (*tbuf.meta));
466 	kmem_free(tbuf.data, tbuf.cnt * sizeof (*tbuf.data));
467 }
468 
469 static inline void
470 xdt_process_rec(uint_t cpuid, struct t_rec *rec)
471 {
472 	xdt_schedinfo_t *sp = &xdt_cpu_schedinfo[cpuid];
473 	int eid;
474 
475 	ASSERT(rec != NULL);
476 	ASSERT(xdt_ncpus == xen_get_nphyscpus());
477 
478 	if (cpuid >= xdt_ncpus) {
479 		tbuf.stat_spurious_cpu++;
480 		return;
481 	}
482 
483 	switch (rec->event) {
484 
485 	/*
486 	 * Sched probes
487 	 */
488 	case TRC_SCHED_SWITCH_INFPREV:
489 		/*
490 		 * Info on vCPU being de-scheduled
491 		 *
492 		 * rec->data[0] = prev domid
493 		 * rec->data[1] = time spent on pcpu
494 		 */
495 		sp->prev_domid = rec->data[0];
496 		sp->prev_ctime = rec->data[1];
497 		break;
498 
499 	case TRC_SCHED_SWITCH_INFNEXT:
500 		/*
501 		 * Info on next vCPU to be scheduled
502 		 *
503 		 * rec->data[0] = next domid
504 		 * rec->data[1] = time spent waiting to get on cpu
505 		 * rec->data[2] = time slice
506 		 */
507 		sp->next_domid = rec->data[0];
508 		sp->next_wtime = rec->data[1];
509 		sp->next_ts = rec->data[2];
510 		break;
511 
512 	case TRC_SCHED_SWITCH:
513 		/*
514 		 * vCPU switch
515 		 *
516 		 * rec->data[0] = prev domid
517 		 * rec->data[1] = prev vcpuid
518 		 * rec->data[2] = next domid
519 		 * rec->data[3] = next vcpuid
520 		 */
521 		if (rec->data[0] != sp->prev_domid &&
522 		    rec->data[2] != sp->next_domid) {
523 			/* prev and next info don't match doms being sched'd */
524 			tbuf.stat_spurious_switch++;
525 			return;
526 		}
527 
528 		sp->prev_vcpuid = rec->data[1];
529 		sp->next_vcpuid = rec->data[3];
530 
531 		XDT_PROBE3(IS_IDLE_DOM(sp->prev_domid)?
532 		    XDT_SCHED_IDLE_OFF_CPU:XDT_SCHED_OFF_CPU,
533 		    cpuid, sp->prev_domid, sp->prev_vcpuid, sp->prev_ctime);
534 
535 		XDT_PROBE4(IS_IDLE_DOM(sp->next_domid)?
536 		    XDT_SCHED_IDLE_ON_CPU:XDT_SCHED_ON_CPU,
537 		    cpuid, sp->next_domid, sp->next_vcpuid, sp->next_wtime,
538 		    sp->next_ts);
539 		break;
540 
541 	case TRC_SCHED_BLOCK:
542 		/*
543 		 * vCPU blocked
544 		 *
545 		 * rec->data[0] = domid
546 		 * rec->data[1] = vcpuid
547 		 */
548 		XDT_PROBE2(XDT_SCHED_BLOCK, cpuid, rec->data[0], rec->data[1]);
549 		break;
550 
551 	case TRC_SCHED_SLEEP:
552 		/*
553 		 * Put vCPU to sleep
554 		 *
555 		 * rec->data[0] = domid
556 		 * rec->data[1] = vcpuid
557 		 */
558 		XDT_PROBE2(XDT_SCHED_SLEEP, cpuid, rec->data[0], rec->data[1]);
559 		break;
560 
561 	case TRC_SCHED_WAKE:
562 		/*
563 		 * Wake up vCPU
564 		 *
565 		 * rec->data[0] = domid
566 		 * rec->data[1] = vcpuid
567 		 */
568 		XDT_PROBE2(XDT_SCHED_WAKE, cpuid, rec->data[0], rec->data[1]);
569 		break;
570 
571 	case TRC_SCHED_YIELD:
572 		/*
573 		 * vCPU yielded
574 		 *
575 		 * rec->data[0] = domid
576 		 * rec->data[1] = vcpuid
577 		 */
578 		XDT_PROBE2(XDT_SCHED_YIELD, cpuid, rec->data[0], rec->data[1]);
579 		break;
580 
581 	case TRC_SCHED_SHUTDOWN:
582 		/*
583 		 * Guest shutting down
584 		 *
585 		 * rec->data[0] = domid
586 		 * rec->data[1] = initiating vcpu
587 		 * rec->data[2] = shutdown code
588 		 */
589 		switch (rec->data[2]) {
590 		case SHUTDOWN_poweroff:
591 			eid = XDT_SCHED_SHUTDOWN_POWEROFF;
592 			break;
593 		case SHUTDOWN_reboot:
594 			eid = XDT_SCHED_SHUTDOWN_REBOOT;
595 			break;
596 		case SHUTDOWN_suspend:
597 			eid = XDT_SCHED_SHUTDOWN_SUSPEND;
598 			break;
599 		case SHUTDOWN_crash:
600 			eid = XDT_SCHED_SHUTDOWN_CRASH;
601 			break;
602 		default:
603 			tbuf.stat_unknown_shutdown++;
604 			return;
605 		}
606 
607 		XDT_PROBE1(eid, cpuid, rec->data[0]);
608 		break;
609 
610 	/*
611 	 * Mem probes
612 	 */
613 	case TRC_MEM_PAGE_GRANT_MAP:
614 		/*
615 		 * Guest mapped page grant
616 		 *
617 		 * rec->data[0] = domid
618 		 */
619 		XDT_PROBE1(XDT_MEM_PAGE_GRANT_MAP, cpuid, rec->data[0]);
620 		break;
621 
622 	case TRC_MEM_PAGE_GRANT_UNMAP:
623 		/*
624 		 * Guest unmapped page grant
625 		 *
626 		 * rec->data[0] = domid
627 		 */
628 		XDT_PROBE1(XDT_MEM_PAGE_GRANT_UNMAP, cpuid, rec->data[0]);
629 		break;
630 
631 	case TRC_MEM_PAGE_GRANT_TRANSFER:
632 		/*
633 		 * Page grant is being transferred
634 		 *
635 		 * rec->data[0] = target domid
636 		 */
637 		XDT_PROBE1(XDT_MEM_PAGE_GRANT_TRANSFER, cpuid, rec->data[0]);
638 		break;
639 
640 	/*
641 	 * HVM probes
642 	 */
643 	case TRC_HVM_VMENTRY:
644 		/*
645 		 * Return to guest via vmx_launch/vmrun
646 		 *
647 		 * rec->data[0] = (domid<<16 + vcpuid)
648 		 */
649 		XDT_PROBE2(XDT_HVM_VMENTRY, cpuid, HVM_DOMID(rec->data[0]),
650 		    HVM_VCPUID(rec->data[0]));
651 		break;
652 
653 	case TRC_HVM_VMEXIT:
654 		/*
655 		 * Entry into VMEXIT handler
656 		 *
657 		 * rec->data[0] = (domid<<16 + vcpuid)
658 		 * rec->data[1] = guest rip
659 		 * rec->data[2] = cpu vendor specific exit code
660 		 */
661 		XDT_PROBE4(XDT_HVM_VMEXIT, cpuid, HVM_DOMID(rec->data[0]),
662 		    HVM_VCPUID(rec->data[0]), rec->data[1], rec->data[2]);
663 		break;
664 
665 	case TRC_LOST_RECORDS:
666 		tbuf.stat_dropped_recs++;
667 		break;
668 
669 	default:
670 		tbuf.stat_unknown_recs++;
671 		break;
672 	}
673 }
674 
675 /*ARGSUSED*/
676 static void
677 xdt_tbuf_scan(void *arg)
678 {
679 	uint_t cpuid;
680 	size_t nrecs;
681 	struct t_rec *rec;
682 	uint32_t prod;
683 
684 	nrecs = (tbuf.size - sizeof (struct t_buf)) / sizeof (struct t_rec);
685 
686 	/* scan all cpu buffers for new records */
687 	for (cpuid = 0; cpuid < tbuf.cnt; cpuid++) {
688 		prod = tbuf.meta[cpuid]->prod;
689 		membar_consumer(); /* read prod /then/ data */
690 		while (tbuf.meta[cpuid]->cons != prod) {
691 			rec = tbuf.data[cpuid] + tbuf.meta[cpuid]->cons % nrecs;
692 			xdt_process_rec(cpuid, rec);
693 			membar_exit(); /* read data /then/ update cons */
694 			tbuf.meta[cpuid]->cons++;
695 		}
696 	}
697 }
698 
699 static void
700 xdt_cyclic_enable(void)
701 {
702 	cyc_handler_t hdlr;
703 	cyc_time_t when;
704 
705 	ASSERT(MUTEX_HELD(&cpu_lock));
706 
707 	hdlr.cyh_func = xdt_tbuf_scan;
708 	hdlr.cyh_arg = NULL;
709 	hdlr.cyh_level = CY_LOW_LEVEL;
710 
711 	when.cyt_interval = xdt_poll_nsec;
712 	when.cyt_when = dtrace_gethrtime() + when.cyt_interval;
713 
714 	xdt_cyclic = cyclic_add(&hdlr, &when);
715 }
716 
717 static void
718 xdt_probe_create(xdt_probe_t *p)
719 {
720 	ASSERT(p != NULL && p->pr_mod != NULL);
721 
722 	if (dtrace_probe_lookup(xdt_id, p->pr_mod, NULL, p->pr_name) != 0)
723 		return;
724 
725 	xdt_prid[p->evt_id] = dtrace_probe_create(xdt_id, p->pr_mod, NULL,
726 	    p->pr_name, dtrace_mach_aframes(), p);
727 }
728 
729 /*ARGSUSED*/
730 static void
731 xdt_provide(void *arg, const dtrace_probedesc_t *desc)
732 {
733 	const char *mod, *name;
734 	int i;
735 
736 	if (desc == NULL) {
737 		for (i = 0; xdt_probe[i].pr_mod != NULL; i++) {
738 			xdt_probe_create(&xdt_probe[i]);
739 		}
740 	} else {
741 		mod = desc->dtpd_mod;
742 		name = desc->dtpd_name;
743 		for (i = 0; xdt_probe[i].pr_mod != NULL; i++) {
744 			int l1 = strlen(xdt_probe[i].pr_name);
745 			int l2 = strlen(xdt_probe[i].pr_mod);
746 			if (strncmp(name, xdt_probe[i].pr_name, l1) == 0 &&
747 			    strncmp(mod, xdt_probe[i].pr_mod, l2) == 0)
748 				break;
749 		}
750 
751 		if (xdt_probe[i].pr_mod == NULL)
752 			return;
753 		xdt_probe_create(&xdt_probe[i]);
754 	}
755 
756 }
757 
758 /*ARGSUSED*/
759 static void
760 xdt_destroy(void *arg, dtrace_id_t id, void *parg)
761 {
762 	xdt_probe_t *p = parg;
763 	xdt_prid[p->evt_id] = 0;
764 }
765 
766 static void
767 xdt_set_trace_mask(uint32_t mask)
768 {
769 	xen_sysctl_tbuf_op_t tbuf_op;
770 
771 	tbuf_op.cmd  = XEN_SYSCTL_TBUFOP_set_evt_mask;
772 	tbuf_op.evt_mask = mask;
773 	(void) xdt_sysctl_tbuf(&tbuf_op);
774 }
775 
776 /*ARGSUSED*/
777 static void
778 xdt_enable(void *arg, dtrace_id_t id, void *parg)
779 {
780 	xdt_probe_t *p = parg;
781 	xen_sysctl_tbuf_op_t tbuf_op;
782 
783 	ASSERT(MUTEX_HELD(&cpu_lock));
784 	ASSERT(xdt_prid[p->evt_id] != 0);
785 
786 	xdt_probemap[p->evt_id] = xdt_prid[p->evt_id];
787 	xdt_classinfo[p->class].cnt++;
788 
789 	if (xdt_classinfo[p->class].cnt == 1) {
790 		/* set the trace mask for this class */
791 		cur_trace_mask |= xdt_classinfo[p->class].trc_mask;
792 		xdt_set_trace_mask(cur_trace_mask);
793 	}
794 
795 	if (xdt_cyclic == CYCLIC_NONE) {
796 		/*
797 		 * DTrace doesn't have the notion of failing an enabling. It
798 		 * works on the premise that, if you have advertised a probe
799 		 * via the pops->dtps_provide() function, you can enable it.
800 		 * Failure is not an option. In the case where we can't enable
801 		 * Xen tracing the consumer will carry on regardless and
802 		 * think all is OK except the probes will never fire.
803 		 */
804 		tbuf_op.cmd = XEN_SYSCTL_TBUFOP_enable;
805 		if (xdt_sysctl_tbuf(&tbuf_op) != 0) {
806 			cmn_err(CE_NOTE, "Couldn't enable hypervisor tracing.");
807 			return;
808 		}
809 
810 		xdt_cyclic_enable();
811 	}
812 }
813 
814 /*ARGSUSED*/
815 static void
816 xdt_disable(void *arg, dtrace_id_t id, void *parg)
817 {
818 	xdt_probe_t *p = parg;
819 	xen_sysctl_tbuf_op_t tbuf_op;
820 	int i, err;
821 
822 	ASSERT(MUTEX_HELD(&cpu_lock));
823 	ASSERT(xdt_probemap[p->evt_id] != 0);
824 	ASSERT(xdt_probemap[p->evt_id] == xdt_prid[p->evt_id]);
825 	ASSERT(xdt_classinfo[p->class].cnt > 0);
826 
827 	/*
828 	 * We could be here in the slight window between the cyclic firing and
829 	 * a call to dtrace_probe() occurring. We need to be careful if we tear
830 	 * down any shared state.
831 	 */
832 
833 	xdt_probemap[p->evt_id] = 0;
834 	xdt_classinfo[p->class].cnt--;
835 
836 	if (xdt_nr_active_probes() == 0) {
837 		cur_trace_mask = 0;
838 
839 		if (xdt_cyclic == CYCLIC_NONE)
840 			return;
841 
842 		/*
843 		 * We will try to disable the trace buffers. If we fail for some
844 		 * reason we will try again, up to a count of XDT_TBUF_RETRY.
845 		 * If we still aren't successful we try to set the trace mask
846 		 * to 0 in order to prevent trace records from being written.
847 		 */
848 		tbuf_op.cmd = XEN_SYSCTL_TBUFOP_disable;
849 		i = 0;
850 		do {
851 			err = xdt_sysctl_tbuf(&tbuf_op);
852 		} while ((err != 0) && (++i < XDT_TBUF_RETRY));
853 
854 		if (err != 0) {
855 			cmn_err(CE_NOTE,
856 			    "Couldn't disable hypervisor tracing.");
857 			xdt_set_trace_mask(0);
858 		} else {
859 			cyclic_remove(xdt_cyclic);
860 			xdt_cyclic = CYCLIC_NONE;
861 			/*
862 			 * We don't bother making the hypercall to set
863 			 * the trace mask, since it will be reset when
864 			 * tracing is re-enabled.
865 			 */
866 		}
867 	} else if (xdt_classinfo[p->class].cnt == 0) {
868 		cur_trace_mask ^= xdt_classinfo[p->class].trc_mask;
869 		/* other probes are enabled, so add the sub-class mask back */
870 		cur_trace_mask |= 0xF000;
871 		xdt_set_trace_mask(cur_trace_mask);
872 	}
873 }
874 
875 static dtrace_pattr_t xdt_attr = {
876 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_PLATFORM },
877 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_PLATFORM },
878 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
879 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_PLATFORM },
880 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_PLATFORM },
881 };
882 
883 static dtrace_pops_t xdt_pops = {
884 	xdt_provide,		/* dtps_provide() */
885 	NULL,			/* dtps_provide_module() */
886 	xdt_enable,		/* dtps_enable() */
887 	xdt_disable,		/* dtps_disable() */
888 	NULL,			/* dtps_suspend() */
889 	NULL,			/* dtps_resume() */
890 	NULL,			/* dtps_getargdesc() */
891 	NULL,			/* dtps_getargval() */
892 	NULL,			/* dtps_usermode() */
893 	xdt_destroy		/* dtps_destroy() */
894 };
895 
896 static int
897 xdt_attach(dev_info_t *devi, ddi_attach_cmd_t cmd)
898 {
899 	int val;
900 
901 	if (!DOMAIN_IS_INITDOMAIN(xen_info))
902 		return (DDI_FAILURE);
903 
904 	switch (cmd) {
905 	case DDI_ATTACH:
906 		break;
907 
908 	case DDI_RESUME:
909 		/*
910 		 * We might support proper suspend/resume in the future, so,
911 		 * return DDI_FAILURE for now.
912 		 */
913 		return (DDI_FAILURE);
914 
915 	default:
916 		return (DDI_FAILURE);
917 	}
918 
919 	xdt_ncpus = xen_get_nphyscpus();
920 	ASSERT(xdt_ncpus > 0);
921 
922 	if (ddi_create_minor_node(devi, "xdt", S_IFCHR, 0, DDI_PSEUDO, 0) ==
923 	    DDI_FAILURE || xdt_attach_trace_buffers() != 0 ||
924 	    dtrace_register("xdt", &xdt_attr, DTRACE_PRIV_KERNEL, NULL,
925 	    &xdt_pops, NULL, &xdt_id) != 0) {
926 		if (tbuf.va != NULL)
927 			xdt_detach_trace_buffers();
928 		ddi_remove_minor_node(devi, NULL);
929 		return (DDI_FAILURE);
930 	}
931 
932 	val = ddi_getprop(DDI_DEV_T_ANY, devi, DDI_PROP_DONTPASS,
933 	    "xdt_poll_nsec", XDT_POLL_DEFAULT);
934 	xdt_poll_nsec = MAX(val, XDT_POLL_MIN);
935 
936 	xdt_cpu_schedinfo = (xdt_schedinfo_t *)kmem_alloc(xdt_ncpus *
937 	    sizeof (xdt_schedinfo_t), KM_SLEEP);
938 	xdt_init_trace_masks();
939 	xdt_kstat_init();
940 
941 	xdt_devi = devi;
942 	ddi_report_dev(devi);
943 	return (DDI_SUCCESS);
944 }
945 
946 static int
947 xdt_detach(dev_info_t *devi, ddi_detach_cmd_t cmd)
948 {
949 	switch (cmd) {
950 	case DDI_DETACH:
951 		break;
952 
953 	case DDI_SUSPEND:
954 		/*
955 		 * We might support proper suspend/resume in the future. So
956 		 * return DDI_FAILURE for now.
957 		 */
958 		return (DDI_FAILURE);
959 
960 	default:
961 		return (DDI_FAILURE);
962 	}
963 
964 	if (dtrace_unregister(xdt_id) != 0)
965 		return (DDI_FAILURE);
966 
967 	xdt_detach_trace_buffers();
968 	kmem_free(xdt_cpu_schedinfo, xdt_ncpus * sizeof (xdt_schedinfo_t));
969 	if (xdt_cyclic != CYCLIC_NONE)
970 		cyclic_remove(xdt_cyclic);
971 	if (xdt_kstats != NULL)
972 		kstat_delete(xdt_kstats);
973 	xdt_devi = (void *)0;
974 	ddi_remove_minor_node(devi, NULL);
975 
976 	return (DDI_SUCCESS);
977 }
978 
979 /*ARGSUSED*/
980 static int
981 xdt_info(dev_info_t *devi, ddi_info_cmd_t infocmd, void *arg, void **result)
982 {
983 	int error;
984 
985 	switch (infocmd) {
986 	case DDI_INFO_DEVT2DEVINFO:
987 		*result = xdt_devi;
988 		error = DDI_SUCCESS;
989 		break;
990 	case DDI_INFO_DEVT2INSTANCE:
991 		*result = (void *)0;
992 		error = DDI_SUCCESS;
993 		break;
994 	default:
995 		error = DDI_FAILURE;
996 	}
997 	return (error);
998 }
999 
1000 static struct cb_ops xdt_cb_ops = {
1001 	nulldev,		/* open(9E) */
1002 	nodev,			/* close(9E) */
1003 	nodev,			/* strategy(9E) */
1004 	nodev,			/* print(9E) */
1005 	nodev,			/* dump(9E) */
1006 	nodev,			/* read(9E) */
1007 	nodev,			/* write(9E) */
1008 	nodev,			/* ioctl(9E) */
1009 	nodev,			/* devmap(9E) */
1010 	nodev,			/* mmap(9E) */
1011 	nodev,			/* segmap(9E) */
1012 	nochpoll,		/* chpoll(9E) */
1013 	ddi_prop_op,		/* prop_op(9E) */
1014 	NULL,			/* streamtab(9S) */
1015 	D_MP | D_64BIT | D_NEW	/* cb_flag */
1016 };
1017 
1018 static struct dev_ops xdt_ops = {
1019 	DEVO_REV,		/* devo_rev */
1020 	0,			/* devo_refcnt */
1021 	xdt_info,		/* getinfo(9E) */
1022 	nulldev,		/* identify(9E) */
1023 	nulldev,		/* probe(9E) */
1024 	xdt_attach,		/* attach(9E) */
1025 	xdt_detach,		/* detach(9E) */
1026 	nulldev,		/* devo_reset */
1027 	&xdt_cb_ops,		/* devo_cb_ops */
1028 	NULL,			/* devo_bus_ops */
1029 	NULL			/* power(9E) */
1030 };
1031 
1032 
1033 static struct modldrv modldrv = {
1034 	&mod_driverops,
1035 	"Hypervisor event tracing",
1036 	&xdt_ops
1037 };
1038 
1039 static struct modlinkage modlinkage = {
1040 	MODREV_1,
1041 	&modldrv,
1042 	NULL
1043 };
1044 
1045 int
1046 _init(void)
1047 {
1048 	return (mod_install(&modlinkage));
1049 }
1050 
1051 int
1052 _fini(void)
1053 {
1054 	return (mod_remove(&modlinkage));
1055 }
1056 
1057 int
1058 _info(struct modinfo *modinfop)
1059 {
1060 	return (mod_info(&modlinkage, modinfop));
1061 }
1062